Data centre architecture is the organised layout of switches, routers, and cables that connects servers and storage so that data can move quickly and reliably. For the 350-601 exam, you must understand the differences between spine-leaf, two-tier, and three-tier designs because Cisco tests whether you can choose the right architecture for a given scenario. These network topologies determine how well a data centre can scale, handle failures, and manage traffic—and the exam expects you to know the trade-offs cold.
Jump to a section
A simple way to picture Data Center Architecture Overview
An airport terminal is a system for moving people from the street to an airplane gate efficiently and without collisions. The terminal has a single main entrance (the border router), a central security checkpoint (the core layer), a corridor with signs to different concourses (the aggregation layer), and multiple gates leading directly to aircraft (the access layer and servers). In a classic three-tier design, passengers first enter the terminal, then pass through security, then walk down the corridor to their concourse, then go to their gate. This works well for medium traffic. However, when the airport is extremely busy—like a major hub during holiday rush—passengers bottleneck at the security checkpoint and the corridor becomes congested. A newer spine-leaf architecture redesigns the terminal: there is no single security line. Instead, many small security stations (leaf switches) connect directly to a central ring of fast-moving shuttles (spine switches). A passenger can arrive at any gate concourse, clear security right there, and then take a shuttle to any other gate. This removes the bottleneck because every leaf switch can reach every other leaf switch through any spine switch in just one hop. The spine-leaf layout scales much better—you simply add more shuttles or more gates—without redesigning the whole terminal.
In a data centre, the airport terminal becomes the network architecture. The passengers are data packets, the gates are servers, and the shuttles are high-speed switch links. Spine-leaf means every server can talk to every other server with the same low latency, no matter where they are physically placed. That is why modern data centres—especially those running Cisco ACI or cloud-scale workloads—use spine-leaf instead of the older three-tier design.
Data centre architecture defines how the physical and logical parts of a network are arranged to move traffic between servers, storage, and the outside world. The three fundamental architectures you need to know for 350-601 are the three-tier (traditional), two-tier (collapsed core), and spine-leaf (modern). Let us examine each one.
Three-Tier Architecture
Three-tier architecture has three distinct layers: the access layer, the aggregation (or distribution) layer, and the core layer. The access layer is where servers physically connect to the network—usually via top-of-rack (ToR) switches that sit at the top of each server rack. The aggregation layer collects traffic from many access switches and applies policies like security access control lists (ACLs) or routing between different parts of the data centre. The core layer provides high-speed connectivity between the aggregation layer and the outside world (the WAN or internet). Traffic in a three-tier design flows up and down through these layers: a server in rack A sends data to its access switch, which forwards it to the aggregation switch, which forwards it to the core, which sends it back down to the destination rack. This creates a hierarchical tree structure. The main advantage is that it is well understood and works for smaller data centres. The big disadvantage is that as the data centre grows, the higher layers become bottlenecks. Also, if any link or switch in the path fails, connectivity is lost. Redundancy (like having two core switches) helps but adds cost and complexity.
Two-Tier Architecture (Collapsed Core)
Two-tier architecture combines the core and aggregation layers into one layer, leaving only the access layer and a combined core/distribution layer. This simplifies the design and reduces latency because traffic takes one fewer hop. It is common in smaller or medium-sized data centres that do not need the massive scale of a spine-leaf design. The two-tier model is still hierarchical: access switches connect to a pair of core switches, and all inter-server traffic must go through those core switches. The limitation is similar to three-tier—the core layer can become oversubscribed—but it is cheaper and simpler to manage. Cisco sometimes calls this a 'collapsed core' design.
Spine-Leaf Architecture
Spine-leaf (also called a Clos topology) eliminates the hierarchical bottleneck by using two layers only: leaf switches and spine switches. Every leaf switch connects to every spine switch, and every spine switch connects to every leaf switch. No leaf switch connects directly to another leaf switch. All traffic between servers attached to different leaf switches must go through a spine switch. Because every leaf is one hop away from every other leaf (through a spine), latency is predictable and very low. If you need more capacity, you add more spine switches (which increases the number of paths) or more leaf switches (which adds more server ports). The fabric scales linearly—double the number of spine switches and you double the bisection bandwidth (the total capacity between any two halves of the network). This design is the foundation of Cisco Application Centric Infrastructure (ACI) and is used in most modern data centres, including those at hyperscalers like Amazon and Google. The trade-off is that you need more cabling (each leaf must connect to each spine), and the configuration must be consistent across all switches.
Key Concepts to Understand
Oversubscription ratio: The ratio of downstream bandwidth (from leaf to servers) to upstream bandwidth (from leaf to spine). For example, if a leaf switch has 48 ports at 10 Gbps connected to servers (480 Gbps total) and four 40 Gbps uplinks to spines (160 Gbps total), the oversubscription ratio is 3:1. This means when all servers try to send traffic to servers on other leaf switches, they compete for the 160 Gbps uplink. Lower ratios mean less contention.
East-west vs. north-south traffic: East-west traffic is between servers inside the data centre (e.g., a database server talking to an application server). North-south traffic is between the data centre and the outside world (e.g., a user accessing a website). Modern applications (like microservices) generate a lot of east-west traffic, which is why spine-leaf is preferred—it handles east-west traffic much better than hierarchical designs.
Redundancy: In three-tier and two-tier designs, you typically deploy pairs of switches at each layer (active/standby or active/active) and use protocols like Spanning Tree Protocol (STP) to prevent loops. Spine-leaf uses equal-cost multipath (ECMP) routing, where traffic is load-balanced across all available spine switches. This gives both redundancy and better utilisation.
Example Scenario
Imagine a data centre with 100 racks, each containing 20 servers. In a three-tier design, each rack has a ToR access switch. Twenty access switches connect to two aggregation switches, and those aggregation switches connect to two core switches. If one aggregation switch fails, 10 racks lose connectivity to the core. In a spine-leaf design, each rack's leaf switch connects to, say, eight spine switches. If one spine switch fails, traffic simply rebalances across the remaining seven. No rack loses full connectivity. That difference is critical for high-availability applications.
The 350-601 exam expects you to compare these architectures on criteria such as scalability, latency, cost, and ease of troubleshooting. You will see multiple-choice questions that give you a set of requirements (e.g., 'a company needs low latency for all server-to-server traffic and expects to double the number of servers every year') and you must select the appropriate architecture. The correct answer is almost always spine-leaf for modern, scalable needs, but you must also know when three-tier or two-tier might still be appropriate (e.g., for a small office data centre with limited budget).
To summarise: spine-leaf is the architecture of choice for 350-601 because it matches Cisco's ACI design philosophy. Two-tier and three-tier are legacy designs that you must understand to answer comparative questions and to recognise why they are being replaced.
Identify Traffic Pattern
Determine whether the data centre workload is primarily east-west (server-to-server) or north-south (in/out to the internet). This decision drives architecture choice. For east-west-heavy workloads like microservices or big data, spine-leaf is preferred. For simple web serving with low east-west traffic, two-tier may suffice.
Count Server Racks and Ports
Count the number of server racks and the number of servers per rack. Each rack will get a leaf (ToR) switch. Estimate the total required server-facing ports. For example, 40 racks with 24 servers each need 40 leaf switches with at least 24 ports each.
Choose Oversubscription Ratio
Decide on an acceptable oversubscription ratio (e.g., 3:1). This determines how many spine switches are needed. If each leaf has 48x10 Gbps server ports and 4x40 Gbps uplinks, the ratio is 480/160 = 3:1. To lower the ratio, add more uplinks (or faster uplinks) by increasing the number of spine switches.
Cable Leaf to Spine in Full Mesh
Physically connect every leaf switch to every spine switch. For 40 leaf switches and 8 spine switches, this requires 40 * 8 = 320 cables. Use fibre optic cables for longer distances. This full mesh is what gives spine-leaf its predictable low latency.
Configure ECMP and Routing Protocol
Enable a routing protocol (e.g., OSPF or BGP) on all leaf and spine switches. Configure ECMP to load balance traffic across all available spine paths. Ensure that the fabric uses equal-cost routes so that traffic is spread evenly. This step is critical for resilience and performance.
Validate and Monitor Fabric
Use Cisco DCNM or CLI commands to verify that all leaf-to-spine links are up, ECMP is distributing traffic, and no loops exist. Monitor the oversubscription ratio over time. If uplink utilisation exceeds 70%, plan to add more spine switches.
An IT professional designing a data centre network for a mid-sized e-commerce company starts by assessing the company's workload and growth plans. The company runs a web application that uses microservices: each service (user authentication, product catalogue, payment processing) communicates with others constantly. This creates a lot of east-west traffic. The IT engineer knows that a three-tier design would force this traffic through the aggregation and core layers, causing latency and potential bottlenecks as the business grows during peak shopping seasons.
The engineer maps out the current server count: 40 racks, each with 24 servers (960 servers total). The company expects to double to 80 racks within two years. Using a spine-leaf architecture, the engineer plans to deploy 40 leaf switches (one per rack) and initially 8 spine switches. Each leaf switch has 48x10 Gbps server-facing ports and 4x40 Gbps uplinks to the spine. The oversubscription ratio is (48 * 10) / (4 * 40) = 480 / 160 = 3:1, which is acceptable for this workload. When the company adds 40 more racks, the engineer will add 40 more leaf switches and, to keep the oversubscription ratio low, will add 4 more spine switches (for a total of 12). The beauty is that no existing cabling or switch configuration needs to change—only the new leaf switches need to be cabled to all spine switches.
The engineer's day-to-day tasks include:
Configuring leaf and spine switches using Cisco NX-OS or ACI.
Monitoring the fabric for link failures or congestion using tools like Cisco Data Center Network Manager (DCNM).
Adjusting the oversubscription ratio by adding spine switches when utilisation crosses 70% on uplinks.
Troubleshooting connectivity issues by checking the fabric's forwarding tables. In spine-leaf, because all paths are equal-cost, the engineer checks ECMP hashing to ensure traffic is balanced.
Upgrading firmware on switches one at a time (spine by spine) without taking the entire fabric offline.
In a real-world meeting, the IT manager asks the engineer to justify the higher cabling cost of spine-leaf versus a cheaper two-tier design. The engineer explains that the initial cabling cost is higher (each leaf needs a cable to every spine), but the operational savings from not having to redesign the network every time they add racks outweighs the upfront cost. The engineer also points out that spine-leaf supports Cisco ACI's policy-based automation, which reduces configuration errors and speeds up deployment of new applications.
Another scenario: the company experiences a failure of two spine switches simultaneously (due to a power issue). In a three-tier design, losing two core switches would have disconnected half the data centre. In spine-leaf, traffic simply re-routes across the remaining six spine switches. The engineer sees alerts in the monitoring dashboard, logs into the switches, verifies that all leaf switches still have connectivity, and schedules a maintenance window to replace the failed spines.
For the 350-601 exam, you are expected to walk through these scenarios and choose the correct architecture based on specific business requirements. The exam will not ask you to configure a spine-leaf fabric, but it will test your understanding of why one architecture is chosen over another and how the components (leaf, spine, oversubscription ratio, ECMP) work together.
The 350-601 exam tests your knowledge of data centre architectures through multiple-choice questions, drag-and-drop ordering tasks, and scenario-based questions. Here is exactly what you need to focus on.
Key Exam Topics:
Spine-leaf vs. three-tier vs. two-tier: You will be asked to select the correct architecture for a given scenario. Look for keywords such as 'low latency,' 'predictable performance,' 'scalability,' or 'east-west traffic.' These point to spine-leaf. If the scenario mentions 'limited budget' or 'small data centre,' consider two-tier.
Oversubscription ratio calculation: You may be given port counts and link speeds on a leaf switch and asked to compute the ratio. Memorise the formula: (total downstream bandwidth) / (total upstream bandwidth).
Role of leaf and spine switches: Leaf switches connect to servers. Spine switches connect only to leaf switches. They do not connect to each other or to servers. That is a common trap: a question might list 'spine switch connects to core switch' or 'spine switch connects to server.' Both are false.
East-west vs. north-south traffic: You must know which architecture handles east-west traffic better and why. A question may ask: 'Which architecture minimises the number of hops for server-to-server traffic?' The answer is spine-leaf (one hop through a spine).
Redundancy mechanisms: In spine-leaf, redundancy comes from ECMP (equal-cost multipath). In three-tier, it comes from STP (Spanning Tree Protocol) and VPC (Virtual Port Channel). The exam may ask which protocol allows load balancing across multiple paths. That is ECMP.
Common Traps:
Confusing the 'core' layer in three-tier with 'spine' in spine-leaf. They are not the same. The core layer is a bottleneck; the spine layer is a fabric that distributes traffic.
Thinking that spine-leaf is always better. The exam will present questions where cost or simplicity is the priority, and two-tier might be the correct choice. Always read the requirements carefully.
Forgetting that leaf switches do not talk directly to each other. Traffic between leafs must go through a spine. A question might imply that leaf switches can be connected directly for efficiency. That is wrong.
Miscomputing oversubscription ratio. The ratio is typically expressed as N:1 (e.g., 3:1). A ratio of 1:1 means no oversubscription—every port can simultaneously communicate at full speed—which is rare and expensive.
Key Definitions to Memorise: - Bisection bandwidth: The minimum bandwidth between two equal halves of the network. In spine-leaf, adding spines increases this. - Collapsed core: Another name for two-tier design where the core and distribution layers are combined. - ToR switch: A leaf switch that sits at the top of a server rack. - ECMP: A routing technique that splits traffic across multiple equal-cost paths; used in spine-leaf fabrics.
Question Types That Appear: - 'Which of the following is an advantage of spine-leaf over three-tier architecture?' Correct answers: lower latency, easier scalability, deterministic paths. Wrong answers: lower cost, simpler cabling, easier to manage (spine-leaf actually requires more cabling and careful configuration). - 'A company requires zero packet loss during a spine switch failure. Which feature is required?' Answer: ECMP with fast convergence. - Drag-and-drop: Place the layers of a three-tier architecture in the correct order (access, aggregation, core from bottom to top).
To prepare, practise comparing architectures side by side. Create a table with columns: Architecture, Latency, Scalability, Cost, Redundancy Method. Fill it in for all three types. This will help you answer the comparative questions quickly.
Spine-leaf architecture uses a full mesh between leaf and spine switches, ensuring every server is exactly one hop away from any other server through the fabric.
Three-tier architecture has three layers (access, aggregation, core) and creates bottlenecks at the aggregation and core layers because all traffic passes through them.
East-west traffic (server-to-server) is the dominant traffic pattern in modern applications, and spine-leaf handles it far more efficiently than hierarchical designs.
Oversubscription ratio = downstream bandwidth / upstream bandwidth; a lower ratio means less contention for uplink capacity.
In spine-leaf, spine switches never connect to servers or to each other; they only connect to leaf switches.
ECMP (Equal-Cost Multipath) is the redundancy mechanism in spine-leaf, replacing Spanning Tree Protocol used in older designs.
Adding spine switches increases bisection bandwidth linearly, making spine-leaf highly scalable for growing data centres.
Two-tier (collapsed core) is appropriate only for small or budget-constrained data centres that do not need high east-west throughput.
Cisco ACI implements spine-leaf topology with policy-based automation for consistent configuration across the fabric.
The 350-601 exam expects you to compare architectures on latency, scalability, cost, and resilience, and to calculate oversubscription ratios.
These come up on the exam all the time. Here's how to tell them apart.
Spine-Leaf
All leaf switches connect to all spine switches in a full mesh.
Latency is consistent and low because any two leafs communicate through only one spine hop.
Scaling is linear: add more leafs or spines without redesigning the network.
Three-Tier
Access switches connect only to one or two aggregation switches, creating a tree.
Latency varies and can be higher because traffic passes through access, aggregation, then core layers.
Scaling often requires upgrading aggregation and core switches, leading to costly forklift upgrades.
Two-Tier (Collapsed Core)
Combines core and distribution into one layer, but still a hierarchy where access switches connect to core switches.
Limited east-west throughput because core switches can become congested.
Cost-effective for small data centres with less than 10 racks.
Spine-Leaf
Two layers with full mesh connectivity—no single bottleneck layer.
High east-west throughput due to multiple parallel spine paths and ECMP load balancing.
Higher initial cabling cost but scales seamlessly to hundreds of racks.
ECMP (Equal-Cost Multipath)
Used in spine-leaf to load balance traffic across multiple equal-cost paths.
Provides both redundancy and increased bandwidth utilisation (all paths active).
Operates at Layer 3 using routing protocols like OSPF or BGP.
STP (Spanning Tree Protocol)
Used in three-tier and two-tier to prevent loops by blocking redundant links.
Only one path is active at a time; other links sit idle until a failover occurs.
Operates at Layer 2; does not utilise available bandwidth efficiently.
Mistake
Spine-leaf and three-tier are the same thing because both have two layers of switches.
Correct
Spine-leaf has only two logical layers (leaf and spine) and every leaf connects to every spine, creating a full mesh. Three-tier has three layers (access, aggregation, core) and uses a tree structure where access switches connect only to one aggregation switch, creating bottlenecks.
Beginners see 'two layers' in spine-leaf and 'three layers' in three-tier and assume that two-tier designs (collapsed core) are the same as spine-leaf. They do not realise the crucial difference in connectivity pattern: full mesh vs. hierarchical.
Mistake
In spine-leaf, leaf switches can connect directly to each other to reduce hops.
Correct
Leaf switches in a spine-leaf fabric never connect directly to each other. All inter-leaf traffic must traverse a spine switch. This rule ensures predictable latency and prevents loops.
It is intuitive to think that connecting switches directly would be faster, but that would create a mesh that is hard to scale and manage. The discipline of always going through the spine is the key to the architecture's simplicity.
Mistake
Oversubscription ratio is a fixed value that should be the same for all data centre designs.
Correct
Oversubscription ratio is a design choice that varies based on workload. Some applications (e.g., high-performance computing) require 1:1 (no oversubscription), while others tolerate 3:1 or higher. Spine-leaf makes it easier to adjust the ratio by adding spines, but three-tier makes it costly to change.
Newcomers often think there is a 'correct' ratio for all scenarios. In reality, the ratio depends on traffic patterns and budget. The exam tests your ability to calculate and interpret the ratio, not just memorise a number.
Mistake
The core layer in three-tier and the spine layer in spine-leaf perform the same function.
Correct
The core layer in three-tier acts as a central aggregation point and can become a bottleneck. The spine layer in spine-leaf is a non-blocking fabric that simply forwards traffic between leafs, and because there are many spines, no single device is a bottleneck.
Both sit in the middle of the topology, so beginners assume they are interchangeable. However, their role in traffic flow and redundancy is fundamentally different. The core is a funnel; the spine is a grid.
Mistake
Two-tier architecture is the same as spine-leaf because both have two layers.
Correct
Two-tier (collapsed core) still uses a hierarchical tree: access switches connect to one or two core switches. It is NOT a full mesh. Spine-leaf uses a leaf-to-spine full mesh. The exam expects you to distinguish them clearly.
The name 'two-tier' sounds like it might be a modern alternative to three-tier, but it is just a simplified version of the legacy design. Spine-leaf is architecturally distinct, not just a smaller version.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
Leaf switches connect to servers (and sometimes to storage). Spine switches connect only to leaf switches. Spine switches do not connect to servers. Traffic between leaf switches must go through a spine switch.
Yes, spine-leaf is a modern implementation of the Clos network topology invented by Charles Clos in the 1950s for telephone switching. The term 'spine-leaf' is used in data centre networking, while 'Clos' is the general mathematical model.
You can, but it is not recommended for large-scale or high-performance environments. Three-tier creates bottlenecks and is harder to scale. For the 350-601 exam, expect questions that highlight why spine-leaf is superior for modern data centres.
Bisection bandwidth is the minimum bandwidth available between two equally sized halves of the network. In spine-leaf, adding more spine switches increases bisection bandwidth, which improves the network's ability to handle high traffic volumes. It is a key measure of scalability.
Oversubscription ratio = total server-facing bandwidth (downstream) divided by total uplink bandwidth (upstream). For example, a leaf with 48 ports of 10 Gbps (480 Gbps down) and 4 ports of 40 Gbps (160 Gbps up) gives a ratio of 480/160 = 3:1.
Spanning Tree Protocol (STP) is used in three-tier and two-tier designs to prevent loops when redundant links are used. In spine-leaf, STP is not needed because the architecture uses Layer 3 routing and ECMP, which inherently avoids loops.
No, the 350-601 exam does not require you to configure a spine-leaf fabric. You need to understand the concepts, compare architectures, and interpret scenarios to choose the correct design. Commands are not tested, but terms like ECMP, leaf, spine, and oversubscription are.
You've finished Data Center Architecture Overview. Continue through the 350-601 study guide to build a complete picture of the exam.
Done with this chapter?